OpenXLA-specific changes

Author: Aliia Khasanova

BUILD

@@ -28,15 +28,16 @@ inline Type bf16Ty(MLIRContext *ctx) { return BFloat16Type::get(ctx); }
 
 inline bool isFloat(Type type) {
   return type.isF32() || type.isF64() || type.isF16() || type.isF128() ||
-         type.isBF16() || type.isFloat8E4M3B11FNUZ() || type.isFloat8E4M3FN() ||
-         type.isFloat8E4M3FNUZ() || type.isFloat8E5M2() ||
-         type.isFloat8E5M2FNUZ();
+         type.isBF16() ||
+         llvm::isa<mlir::Float8E4M3B11FNUZType, mlir::Float8E4M3FNType,
+                   mlir::Float8E4M3FNUZType, mlir::Float8E5M2Type,
+                   mlir::Float8E5M2FNUZType>(type);
 }
 
 inline bool isFloat8(Type type) {
-  return type.isFloat8E4M3B11FNUZ() || type.isFloat8E4M3FN() ||
-         type.isFloat8E4M3FNUZ() || type.isFloat8E5M2() ||
-         type.isFloat8E5M2FNUZ();
+  return llvm::isa<mlir::Float8E4M3B11FNUZType, mlir::Float8E4M3FNType,
+                   mlir::Float8E4M3FNUZType, mlir::Float8E5M2Type,
+                   mlir::Float8E5M2FNUZType>(type);
 }
 
 inline bool isInt(Type type) { return type.isIntOrFloat() && !isFloat(type); }

include/triton/Conversion/MLIRTypes.h

@@ -1105,7 +1105,12 @@ def CallOp : TT_Op<"call", [CallOpInterface, /*MemRefsNormalizable, */DeclareOpI
     MutableOperandRange getArgOperandsMutable() {
       return getOperandsMutable();
     }
-
+    Attribute removeArgAttrsAttr() { return nullptr; }
+    Attribute removeResAttrsAttr() { return nullptr; }
+    ArrayAttr getArgAttrsAttr() { return nullptr; }
+    ArrayAttr getResAttrsAttr() { return nullptr; }
+    void setArgAttrsAttr(ArrayAttr) { return; }
+    void setResAttrsAttr(ArrayAttr) { return; }
   }];
 
   let assemblyFormat = [{

include/triton/Dialect/Triton/IR/TritonOps.td

@@ -123,6 +123,12 @@ ScratchConfig getScratchConfigForCvt(RankedTensorType srcTy,
 
   std::tie(scratchConfig.inVec, scratchConfig.outVec) =
       getScratchCvtInOutVecLengths(srcTy, dstTy);
+  // We can't write a longer vector than the shape of shared memory.
+  // This shape might be smaller than the tensor shape in case we decided to
+  // do the conversion in multiple iterations.
+  unsigned contiguousShapeDim = scratchConfig.repShape[scratchConfig.order[0]];
+  scratchConfig.inVec = std::min(scratchConfig.inVec, contiguousShapeDim);
+  scratchConfig.outVec = std::min(scratchConfig.outVec, contiguousShapeDim);
 
   // No padding is required if the tensor is 1-D, or if all dimensions except
   // the first accessed dimension have a size of 1.

lib/Analysis/Allocation.cpp

@@ -935,7 +935,7 @@ class ShROpAxisInfoVisitor final : public BinaryOpVisitorImpl<OpTy> {
       // Treat [2^n,2^n+1,...]'s divisibility as 1 instead of 2^n
       lhsDivisibility = 1;
     }
-    return std::max<int64_t>(1, lhsDivisibility / (1 << shift));
+    return std::max<int64_t>(1, lhsDivisibility / (int64_t(1) << shift));
   }
 
   int64_t getConstancy(OpTy op, const AxisInfo &lhs, const AxisInfo &rhs,

lib/Analysis/AxisInfo.cpp

@@ -750,14 +750,14 @@ bool supportMMA(triton::DotOp op, int version) {
       return false;
     if (!(numWarps % 4 == 0 && retShapePerCTA[rank - 2] % 64 == 0 &&
           retShapePerCTA[rank - 1] % 8 == 0 &&
-          (aElemTy.isFloat8E5M2() || aElemTy.isFloat8E4M3FN() ||
+          (llvm::isa<mlir::Float8E5M2Type, mlir::Float8E4M3FNType>(aElemTy) ||
            aElemTy.isInteger(8) || aElemTy.isF16() || aElemTy.isBF16() ||
            aElemTy.isF32()))) {
       return false;
     }
     // We cannot use MMA_V3 if we need to accumulate in F32 within the MMA op.
     if (op.getMaxNumImpreciseAcc() < 32 &&
-        (aElemTy.isFloat8E5M2() || aElemTy.isFloat8E4M3FN()) &&
+        (llvm::isa<mlir::Float8E5M2Type, mlir::Float8E4M3FNType>(aElemTy)) &&
         cast<RankedTensorType>(op.getType()).getElementType().isF32()) {
       return false;
     }
@@ -778,8 +778,9 @@ bool supportMMA(Value value, int version) {
       cast<triton::gpu::TensorOrMemDesc>(value.getType()).getElementType();
   // FP8 is not natively supported on all mma versions but it can always be
   // promoted to fp16 therefore we can always support it.
-  bool isFP8 = elemTy.isFloat8E5M2() || elemTy.isFloat8E4M3FN() ||
-               elemTy.isFloat8E5M2FNUZ() || elemTy.isFloat8E4M3FNUZ();
+  bool isFP8 =
+      llvm::isa<mlir::Float8E5M2Type, mlir::Float8E4M3FNType,
+                mlir::Float8E5M2FNUZType, mlir::Float8E4M3FNUZType>(elemTy);
   return isFP8 || elemTy.isF16() || elemTy.isBF16() ||
          (elemTy.isF32() && version >= 2) ||
          (elemTy.isInteger(8) && version >= 2);

lib/Analysis/Utility.cpp

@@ -57,6 +57,12 @@ TritonGPUTypeConverter::TritonGPUTypeConverter(MLIRContext *context,
   addArgumentMaterialization([&](OpBuilder &builder,
                                  RankedTensorType tensorType, ValueRange inputs,
                                  Location loc) -> Value {
+    // Allows partial TTIR to TTGIR conversion by materializing a conversion for
+    // remaining arguments that have been converted to a new type.
+    // We use this to rewrite triton_xla.sparse_dot in a separate pass after
+    // 'convert-triton-to-tritongpu'.
+    return builder.create<triton::gpu::ConvertLayoutOp>(loc, tensorType,
+                                                        inputs);
     llvm_unreachable("Argument rematerialization should not happen in Triton "
                      "-> TritonGPU conversion");
     return {};
@@ -66,6 +72,12 @@ TritonGPUTypeConverter::TritonGPUTypeConverter(MLIRContext *context,
   // convert origValue to newValue
   addSourceMaterialization([&](OpBuilder &builder, RankedTensorType tensorType,
                                ValueRange inputs, Location loc) -> Value {
+    // Allows partial TTIR to TTGIR conversion by materializing a conversion for
+    // remaining uses of values that have been converted to a new type.
+    // We use this to rewrite triton_xla.sparse_dot in a separate pass after
+    // 'convert-triton-to-tritongpu'.
+    return builder.create<triton::gpu::ConvertLayoutOp>(loc, tensorType,
+                                                        inputs);
     llvm_unreachable("Source rematerialization should not happen in Triton -> "
                      "TritonGPU Conversion");
     return {};

lib/Conversion/TritonToTritonGPU/TritonGPUConversion.cpp

@@ -899,7 +899,7 @@ void FuncOp::build(OpBuilder &builder, OperationState &state, StringRef name,
   if (argAttrs.empty())
     return;
   assert(type.getNumInputs() == argAttrs.size());
-  function_interface_impl::addArgAndResultAttrs(
+  call_interface_impl::addArgAndResultAttrs(
       builder, state, argAttrs, /*resultAttrs=*/std::nullopt,
       getArgAttrsAttrName(state.name), getResAttrsAttrName(state.name));
 }

lib/Dialect/Triton/IR/Ops.cpp

@@ -151,6 +151,11 @@ struct CanonicalizeConvertFromAlloc
     auto convert = op.getSrc().getDefiningOp<ConvertLayoutOp>();
     if (!convert)
       return failure();
+    // LocalAllocOp lowering doesn't support going from DotOperandEncoding
+    // to SharedEncoding, so we want to keep this layout conversion.
+    if (mlir::isa<triton::gpu::DotOperandEncodingAttr>(
+            convert.getSrc().getType().getEncoding()))
+      return failure();
     rewriter.replaceOpWithNewOp<triton::gpu::LocalAllocOp>(
         op, op->getResult(0).getType(), convert.getSrc());
     return mlir::success();
@@ -213,13 +218,13 @@ struct CanonicalizeConvertFromConvert
     // heuristic to accommodate fused attention.
     auto srcType = op.getSrc().getType();
     auto dstType = op.getType();
-    if (mlir::isa<DotOperandEncodingAttr>(dstType.getEncoding()) &&
-        mlir::isa<NvidiaMmaEncodingAttr>(srcType.getEncoding()))
+    if (mlir::isa_and_nonnull<DotOperandEncodingAttr>(dstType.getEncoding()) &&
+        mlir::isa_and_nonnull<NvidiaMmaEncodingAttr>(srcType.getEncoding()))
       return failure();
 
     // for hopper MMAv3
-    if (mlir::isa<SharedEncodingAttr>(dstType.getEncoding()) &&
-        mlir::isa<NvidiaMmaEncodingAttr>(srcType.getEncoding()) &&
+    if (mlir::isa_and_nonnull<SharedEncodingAttr>(dstType.getEncoding()) &&
+        mlir::isa_and_nonnull<NvidiaMmaEncodingAttr>(srcType.getEncoding()) &&
         llvm::any_of(op.getResult().getUsers(), [](Operation *dot) {
           return dot->hasTrait<OpTrait::DotLike>();
         })) {

lib/Dialect/TritonGPU/IR/Ops.cpp

@@ -21,8 +21,6 @@ namespace mlir {
 namespace triton {
 namespace gpu {
 
-namespace {
-
 // Get the highest version supported for the hardware and the dot.
 static int getMMAVersionSafe(int computeCapability, DotOp op) {
   // List supported mma version in order of preference.
@@ -47,8 +45,8 @@ static int getMMAVersionSafe(int computeCapability, DotOp op) {
   return 0;
 }
 
-SmallVector<unsigned> warpsPerTileV2(DotOp dotOp, const ArrayRef<int64_t> shape,
-                                     int numWarps) {
+SmallVector<unsigned>
+warpsPerTileV2(Operation *dotOp, const ArrayRef<int64_t> shape, int numWarps) {
   auto rank = shape.size();
   // Early exit for batched matmul
   if (rank == 3)
@@ -112,10 +110,10 @@ SmallVector<unsigned> warpsPerTileV2(DotOp dotOp, const ArrayRef<int64_t> shape,
 }
 
 SmallVector<unsigned, 2>
-warpsPerTileV3(DotOp dotOp, const ArrayRef<int64_t> shape, int numWarps,
+warpsPerTileV3(Operation *dotOp, const ArrayRef<int64_t> shape, int numWarps,
                const SmallVector<unsigned, 3> &instrShape) {
   SetVector<Operation *> slices;
-  mlir::getForwardSlice(dotOp.getResult(), &slices);
+  mlir::getForwardSlice(dotOp->getResult(0), &slices);
   // Contains a chained dot. We prefer to assign warps to one axis
   // to facilitate use cases like flash attention, allowing reductions within
   // the same warp.
@@ -170,11 +168,26 @@ static Value getSharedMemoryMMAOperand(Value v, mlir::PatternRewriter &rewriter,
   auto newType = MemDescType::get(argType.getShape(), argType.getElementType(),
                                   newLayout, SharedMemorySpace);
   rewriter.setInsertionPointAfterValue(arg);
+
+  // LocalAllocOp lowering doesn't support going from DotOperandEncoding
+  // to SharedEncoding.
+  if (auto dotOpEnc = mlir::dyn_cast<DotOperandEncodingAttr>(
+          argType.getEncoding())) {
+    // Create a layout conversion from DotOperandEncoding to BlockedEncoding
+    // then pass it to the LocalAllocOp.
+    auto newArgType = RankedTensorType::get(
+        argType.getShape(), argType.getElementType(), dotOpEnc.getParent());
+    auto dotOperandToBlockedCvt =
+        rewriter.create<ConvertLayoutOp>(arg.getLoc(), newArgType, arg);
+    return rewriter.create<LocalAllocOp>(arg.getLoc(), newType,
+                                              dotOperandToBlockedCvt);
+  }
+
   return rewriter.create<LocalAllocOp>(arg.getLoc(), newType, arg);
 }
 
 SmallVector<unsigned, 3>
-getWarpsPerTile(DotOp dotOp, const ArrayRef<int64_t> shape, int version,
+getWarpsPerTile(Operation* dotOp, const ArrayRef<int64_t> shape, int version,
                 int numWarps, const SmallVector<unsigned, 3> &instrShape) {
   switch (version) {
   case 2:
@@ -188,6 +201,16 @@ getWarpsPerTile(DotOp dotOp, const ArrayRef<int64_t> shape, int version,
 }
 
 static bool bwdFilter(Operation *op) {
+  // Dot operand layout assignment to Predicates are not currently supported
+  // during lowering from TritonGPU to LLVM in Triton for MMA cases. This
+  // condition limits visibility of the original bit-width so that predicate
+  // are not considered, hence, kwidth can never be = 32.
+  if (isa<arith::UIToFPOp>(op)) {
+    Type srcType = getElementTypeOrSelf(op->getOperand(0));
+    if (srcType.isInteger(1))
+      return false;
+  }
+
   return op->getNumOperands() == 1 &&
          (isa<FpToFpOp, BitcastOp, ConvertLayoutOp>(op) ||
           isPureUnaryInlineAsm(op) ||
@@ -207,7 +230,7 @@ static bool bwdFilter(Operation *op) {
 // result, kwidth can be the bitwidth of the lower precision primitive.
 // Conversely, in the downcasting scenario, no reordering is performed,
 // making it directory use the lower precision primitive.
-static int computeOrigBitWidth(Value x) {
+int computeOrigBitWidth(Value x) {
   int finalBitWidth = getElementTypeOrSelf(x).getIntOrFloatBitWidth();
   int origBitWidth = finalBitWidth;
   SetVector<Operation *> slice;
@@ -227,6 +250,9 @@ static int computeOrigBitWidth(Value x) {
   }
   return origBitWidth;
 }
+// Move anonymous namespace down, so getWarpsPerTile is visible to the sparsity
+// extension.
+namespace {
 
 class BlockedToMMA : public mlir::OpRewritePattern<DotOp> {
   int computeCapability;
@@ -632,7 +658,8 @@ static void decomposeMixedModeDotOp(ModuleOp mod, int computeCapability) {
     NvidiaMmaEncodingAttr mmaLayout =
         dyn_cast<NvidiaMmaEncodingAttr>(D.getType().getEncoding());
     if (mmaLayout) {
-      bool isNativeFP8 = AElType.isFloat8E5M2() || AElType.isFloat8E4M3FN();
+      bool isNativeFP8 =
+          llvm::isa<mlir::Float8E5M2Type, mlir::Float8E4M3FNType>(AElType);
       // promote operands for sm < 89 since fp8 mma is not natively supported
       // promote operands for sm >= 90 when mma is not v3
       if (!isNativeFP8 ||
@@ -1018,6 +1045,11 @@ class TritonGPUAccelerateMatmulPass
   }
 };
 
+Value getSharedMemMMAOperand(Value v, mlir::PatternRewriter &rewriter,
+                                int opIdx, bool allowTranspose) {
+  return getSharedMemoryMMAOperand(v, rewriter, opIdx, allowTranspose);
+}
+
 } // namespace gpu
 } // namespace triton
 } // namespace mlir